Documentation · Version 2.3 · Updated July 2026
The Oracle Method: the full specification
The Oracle Method is how Botanik measures a brand's visibility in generative engines, and this page publishes the entire specification: the prompt corpus, the engines queried, the scoring, the handling of instability, and the limits. Version 2.3, updated in July 2026, is free to reuse under CC BY 4.0.
What the Oracle Method measures
The Oracle Method is Botanik's published procedure for measuring whether generative engines cite a brand, built on up to 5,000 demand-weighted prompts replayed through ChatGPT, Perplexity and Google AI. Version 2.3, updated July 2026, applies to the Oracle audit, the Baromètre GEO France (Botanik's public barometer) and the free scan.
This page is the complete, public version of the specification: nothing held back, limits included. It covers, in order: the corpus of prompts, the engines and the conditions under which they are queried, the scoring that turns answers into a comparable number, the treatment of instability, and what the method cannot do.
Botanik publishes it because a number without a method is an opinion. With the specification in the open, anyone can check what a score means, replay the reasoning behind a threshold, or challenge it. The same specification applies to the Oracle audit, to the Baromètre GEO France, and to the free scan.
The key parameters, at a glance. Corpus: up to 5,000 prompts weighted by demand. Engines: ChatGPT, Perplexity and Google AI, measured separately. Runs: 5 per prompt and per engine, with the median reported. Scoring: Reciprocal Rank Fusion with k = 60. Readout: three states per prompt and per engine, Absent, Mentioned or Recommended.
The corpus: up to 5,000 prompts weighted by real demand
A measurement is only as good as the questions asked. The Oracle corpus is built per market: up to 5,000 prompts, each starting as a commercial query extracted from demand data, search and conversational, rewritten as a natural prompt, then weighted by real demand volume.
Weighting is the point. A heavily demanded prompt weighs more than a niche prompt, so the score reflects the questions your market actually asks rather than the questions that are easy to track. The corpus is not a keyword list ported into a chatbot: it is rebuilt for each market from that market's own demand data.
Every prompt starts as a commercial query, extracted from search and conversational demand data, then rewritten as a natural question, the way a person phrases it in a chat window. The rewriting matters because buyers talk to engines in sentences, not in keyword strings, and the measurement has to match real usage.
The corpus spans four intents. Discovery: "how do I choose…". Comparison: "which is the best…". Local: "… near me". Validation: "reviews of…". Each intent captures a different moment in the buying journey, and every prompt is scored the same way regardless of its intent.
The pillars: corpus, engines, scoring, readout
Prompts reflect demand, not a keyword list. The corpus is rebuilt per market from real demand data, then weighted: a heavily demanded prompt weighs more than a niche prompt.
ChatGPT, Perplexity and Google AI receive the same prompts, under the same conditions, via API. They differ in what they cite and how they cite it, so each engine is measured separately.
Every answer is converted into a rank, then fused across prompts and runs, engine by engine, by Reciprocal Rank Fusion. One isolated first place is never enough to dominate the score.
Absent, Mentioned or Recommended: three named states instead of an opaque composite score. You see where you stand, prompt by prompt and engine by engine.
The Oracle Method rests on a small set of design choices: a corpus weighted by real demand, three engines queried under equal conditions, a rank fusion that resists outliers, and a readout in three states. Each choice keeps the measurement honest, comparable, and readable without training.
How the Reciprocal Rank Fusion scoring works
The Oracle Method converts every engine answer into a rank: the brand's position in the answer, where 1 means cited first. Ranks are then fused, engine by engine, across all prompts and all runs with Reciprocal Rank Fusion: score(brand) = Σ 1 / (k + rank), with k = 60.
The sum runs over every prompt and every run. An absence counts as an infinite rank and contributes zero, without breaking the sum. A single isolated first place is not enough to dominate. And because every answer is reduced to a rank, the score does not depend on how an engine formats its answers.
A worked example, with k = 60. A prompt where the brand ranks first contributes 1 / (60 + 1), or 0.0164. A prompt where it ranks fourth contributes 1 / (60 + 4), or 0.0156. A prompt where it is absent contributes 0. An absence forfeits the whole contribution, while the gap between rank 1 and rank 4 stays deliberately small.
That behavior is why rank fusion was chosen over a simple average of positions: it is robust to outliers, it handles absences cleanly, and it rewards consistent presence across many prompts over an occasional spike on one.
Instability: 5 runs, a median, a ± band
Generative engines are not deterministic: the same question can produce different answers. The Oracle Method therefore runs each prompt 5 times per engine, reports the median of the scores, and attaches a ± stability band, the interquartile range, so every number carries its own uncertainty.
The median is reported rather than the mean because a single outlier run should not move the headline number. The ± band shows how much the 5 runs disagreed with each other, so a stable score and a volatile score never look the same on paper.
The reading rule is written into every report: a variation from one edition to the next is significant only if it exceeds the sum of the two stability bands. Below that, the movement is noise, and Botanik writes it as noise, in those words.
Absent, Mentioned, Recommended: the exact thresholds
The three states of the Oracle Method are not impressions. They are thresholds measured on the runs of each prompt: Absent below 5% of runs, Mentioned at 5% or more without being first, Recommended when the brand is the first recommendation in at least 25% of runs. Version 2.3 fixes these thresholds in the published specification.
| State | Threshold | How to read it | Horizon de résultat |
|---|---|---|---|
| Absent | Cited in less than 5% of the prompt's runs | For the engines, the brand does not exist on this question. | |
| Mentioned | Cited in at least 5% of runs, without being the first recommendation | The brand appears in answers but does not lead them. | |
| Recommended | The first recommendation in at least 25% of the prompt's runs | The brand is what the engine puts forward first on this question. |
What this method does not do
5 runs reduce the noise, they do not cancel it. That is what the ± stability band is for, and why it sits next to every score.
An engine can cite an old state of your market, frozen at training time. Every collection is dated for exactly that reason.
French prompts dominate the corpus. Multilingual coverage is partial, and the specification says so rather than implying global reach.
Model updates can shift answers overnight. Comparisons only hold at constant methodology, which is why the method carries a version number.
An honest method documents its blind spots. The Oracle Method carries four known limits, printed as such in every report: residual non-determinism, training windows, a French-dominant corpus, and engines that change without notice. None of them is hidden, and each one shapes how the numbers should be read.
One method, three products
One method, three products at Botanik: the Oracle audit (from €2,670), the Baromètre GEO France and the free scan all apply the Oracle Method, version 2.3. Each product reports the same three-state verdict, Absent, Mentioned or Recommended, engine by engine; what changes is depth and format.
The Oracle audit applies the full specification to one brand on its market, from €2,670 one-shot. The Baromètre GEO France is the public application of the method to French markets, published for anyone to read. The free scan gives a first verdict in 60 seconds, without an account and without a card.
Whatever the product, the readout is the same: a verdict per engine, Absent, Mentioned or Recommended. If you want the method applied to your own brand, the scan is the fastest way to see it working: run it, read the verdict engine by engine, then decide what to do with the result.
How to cite the Oracle Method
The Oracle Method specification is published under CC BY 4.0: reuse is free with attribution. The published citation line reads: "Méthode Oracle v2.3", Botanik, juillet 2026, botanik.ai/methode-oracle. This page is the English version of that specification and carries the same version number, 2.3, and the same license.
If you quote a figure from this page, keep the version number attached. Comparisons across editions only hold at constant methodology, and version 2.3 is the edition updated in July 2026. When the method changes, the number changes with it, and so does the citation line.
Attribution is the only condition. You can reproduce the thresholds, the formula and the limits in your own material, as long as Botanik is credited: that is what CC BY 4.0, free reuse with attribution, means.
How many prompts does the Oracle Method use?
Up to 5,000 prompts per market, weighted by real demand volume. Prompts are commercial queries extracted from search and conversational demand data, rewritten as natural questions, and split across four intents: discovery, comparison, local and validation.
Which engines does the Oracle Method query?
ChatGPT, Perplexity and Google AI, queried under equal conditions: same prompts, same conditions, via API. Each engine is measured separately, because engines differ in what they cite and how they cite it.
How many times is each prompt run?
5 times per prompt and per engine. The method reports the median score with a ± stability band, the interquartile range, because generative engines are not deterministic and a single run proves nothing.
What does Recommended mean in the Oracle Method?
In the Oracle Method, Recommended means the brand is the first recommendation in at least 25% of the runs of a given prompt, measured engine by engine. Mentioned means it appears in at least 5% of runs without being first. Absent means it is cited in less than 5% of runs.
Can I reuse the Oracle Method?
Yes, under CC BY 4.0: reuse is free with attribution. The published citation line is: "Méthode Oracle v2.3", Botanik, juillet 2026, botanik.ai/methode-oracle. This page is the English version of that specification.
How much does the Oracle audit cost?
The Oracle audit starts at €2,670 one-shot. It applies this exact specification, version 2.3, to one brand on its market, engine by engine and prompt by prompt.
Thirty minutes with a consultant
Not a salesperson. We tell you which lever fits your case, and if none does, we say that too.